ESG-Ratings and Returns

(1)

GRA 19703

Master Thesis

Thesis Master of Science

Navn: Dan Fredrik Clemp Ottesen, Sarunas Zilinskas

Start: 15.01.2020 09.00 Finish: 01.09.2020 12.00

(2)

ESG-Ratings and Returns

Master Thesis

by

Ottesen, Dan and Zilinskas, Sarunas MSc in Business with Major Finance

Oslo, August 31, 2020

ABSTRACT

We find that value-weighted portfolios long US stocks from companies with low Environmental, Social and Governance (ESG)-ratings and short stocks with high ESG-ratings have returned annualized 5-factor alphas between 6.9% and 10.8% in the period of 2010 and 2018 depending on the choice of breakpoint. Through analysing holdings of institutional investors, we find that the difference in performance cannot be attributed to behavioral changes such as negative screening of low-rated ESG stocks or impact investing in high-rated ESG stocks.

Supervisor: Patrick Konermann Department of Finance

This thesis is a part of the MSc programme at BI Norwegian Business School. The school takes no responsibility for the methods used, results found,

or conclusions drawn.

(3)

Acknowledgements

Thanks to our supervisor Associate Professor Patrick Konermann who was able to supervise us regardless of the challenges caused due to Covid-19. Thanks to BI for having the foresight to postpone the submission deadline. Thanks to Wharton Research Data Services (WRDS) for publishing brilliant and neat example code that has helped us massively in progressing our coding abilities.

(4)

List of Abbreviations

AMEX American Stock Exchange

CRSP Center for Research in Security Prices DJIA Dow Jones Industrial Average

ESG Environmental, Social and Governance EW Equal-Weighted

IBES Institutional Brokers Estimate Systems KFDL Kenneth French’s Data Library

NYSE New York Stock Exchange VW Value-Weighted

WRDS Wharton Research Data Services

(6)

List of Tables

1 Descriptive Statistics and Correlations Time-Series Regressions

2010-2018 . . . 15

2 Time-Series Regressions - 10^th percentile . . . 16

3 Fama Macbeth - Company, 10% ESG 2010-2018 . . . 18

4 Institutional Ownership, 10% Lowest Rated . . . 20

5 Institutional Ownership Industry Regressions: Low-Rated . . . . 22

6 Industry Time-Series Regressions: Low ESG . . . 24

7 Variable Descriptions and Construction . . . 31

8 Average ESG-Rating, Industry-level . . . 35

9 ESG-Rating Coverage 2003-2018 . . . 36

10 List of Lowest Decile ESG-Companies . . . 37

11 Descriptive Statistics and Correlations Time-Series Regressions 2004-2018 . . . 42

12 Descriptive Statistics Cross-sectional Returns Regressions . . . . 43

13 Correlations Cross-sectional Returns Regressions . . . 44

14 Descriptive Statistics Institutional Ownership-Regressions . . . . 45

15 Correlations Institutional Ownership Regressions . . . 46

16 Time-Series Regressions - Additional Portfolios, 10% . . . 47

17 Fama Macbeth - Additional Regressions . . . 48

18 Institutional Ownership, 10% Highest Rated . . . 49

19 Institutional Ownership Industry Regressions: 1980-2009 . . . . 50

20 Institutional Ownership Industry Regressions: Ex Sin Stocks . . 51

21 Time-Series Regressions - Different Breakpoints . . . 52

22 Time-Series Regressions - Portfolio Formation Timing, 10% . . . 53

23 Time-Series Regressions - Winsorization Levels . . . 54

24 Time-Series Regressions - Ex Announcement Returns . . . 55

25 Fama Macbeth - Ex Announcement Returns . . . 56

26 Institutional Ownership Breakpoints . . . 57

27 Breadth of Ownership, 2010-2018 . . . 58

(7)

List of Symbols

α Parameter for intercept in ESG Portfolios β Factor loading in ESG Portfolios

r_f Parameter for risk-free returns r_i Parameter for returns of stock i r_mkt Parameter for market returns

rESG_low Parameter for stock returns of low-rated ESG companies rESGhigh Parameter for stock returns of high-rated ESG companies rCOM P_low Parameter for stock returns of comparable industries Parameter for error term

a Parameter for intercept in Cross-sectional Regressions

b Coefficient for independent variable in Cross-sectional Regressions c Parameter for intercept in Institutional Ownership Regressions

d Coefficient for independent variable in Institutional Ownership Regressions T Parameter for number of time periods

ρ Parameter for correlation

N_g Parameter for within-cluster correlation τ Parameter for Moulton errors

(8)

1 Introduction

ESG-investment has seen a recent explosion in fund allocation, quadru- pling in size from $3 trillion to $12 trillion between 2010 and 2018 (USSIF (2018)). Even though total funds allocated to ESG investing has been trend- ing sharply upwards, there is little evidence of ESG-based investment strategies causing abnormal, positive returns for investors. The Dow Jones Sustainabil- ity Index North America, a value-weighted index consisting of the top quantile ESG-rated companies out of the largest 600 companies in North America, has under-performed the S&P500 by 2%, and the Dow Jones Industrial Aver- age (DJIA) by 0.7% on an annualized basis since 2010.¹ Even though many investors primarily use ESG information as a risk-assessment tool, some investment strategies are based on invoking behavioral changes in the corporate governance of companies through diverting investment away from irresponsible companies’ stocks, popularly referred to as negative screening, and impact investing, which is to allocate more investment into stocks from companies that are considered positive for society.

Through our thesis we first investigate whether the under-performance of high-rated ESG stocks holds in general, using Fama and French’s 5-factor model (2015) as our primary benchmark. We draw inspiration from previous results presented by Hong and Kaperczyk (2009), who used Merton’s (1987) theories on neglected stocks and segmented markets to show that stocks from the alcohol-, tobacco-, and gambling-industries systematically outperformed a portfolio consisting of stocks from comparable industries with annualized 4-factor alphas of 3.7% between 1926 and 2006. Hong and Kaperczyk hypothesized that institutional investors like pension funds, universities, religious organizations, banks and insurance companies are subject to social norm pressure and therefore are likely to perform negative screening of ’sin’ stocks, leading

1See DJIA North America Composite Index Ticker: A1SGI

(9)

to a synthetic downwards shift in demand compared to their ’non-sinful’ coun- terparts, with the alternative explanation offered that the ’sin’ effect simply is compensation for regulation risk. We investigate whether this effect can be ex- trapolated to the case of portfolios of stocks from companies with high and low ESG-ratings. Our investigations are based on the assumption that stocks from companies with low ESG-ratings should be subject to a decrease in demand due to social norms or increased regulation risk, leading to positive, abnormal returns. Conversely, stocks from companies with high ESG-ratings should under-perform due to increased investment allocation from socially responsible investors.

Our motivation to delve deeper into this topic stems from previous research having mixed evidence on investments strategies that are based on ESG-ratings, the topic of ESG-based investing being a relative new one, and the potential contribution towards better understanding of non-fundamental financial factors’ impact in asset pricing. 2010 serves as a natural starting point for our analysis, not only because of the asset allocation boom, but because 2010 marks the year when our sample of ESG-ratings reaches 10% of total companies listed.

2 Literature Review

Most literature on ESG investing have hypothesized that ESG at its core is an intangible asset, to which investors under-react and therefore a long-term strategy of investing in high-rated ESG assets should yield abnormal returns unexplained by financial factors. Similar under-reaction phenomena that has shown results that cannot be explained by market efficiency theory include post-earnings announcement drift (Bernard and Thomas (1989)) and momentum strategies (Jegadeesh and Titman (1993)). Examples of undervaluation of intangibles include Chan, Lakonishok and Sougiannis (2001), who found that R&D and advertisement intensive firms earned abnormal returns from 1975 to

(10)

1995, and theorized that this may have been a result of accounting rules al- lowing these investments to be expensed rather than put on the balance sheet as an intangible asset. In a similar vein of research, Edmans (2011) found that companies with high employee satisfaction returned an annualized four-factor alpha of 3.5% between 1984 to 2009.

Other reasons for deviations in returns from investing using ESG-based strategies are theories of the downsides of active investing presented by Sharpe (1991), later coined as ’equilibrium accounting’ by Fama and French (2007), who expanded the argument to look at asset prices when a subset of investors treat assets as consumption goods. According to Fama and French, traditional asset pricing models fails to explain behavioral differences that are not rooted in the risk-return relationship of assets and theorize that if a substantial group of investors invest based on non-financial factors, this may pivot the true tan- gency portfolio away from the market-portfolio and make prices become less rational. In a similar vein of research Merton (1987) argues that if certain firms are neglected by investors, these firms’ stocks have a smaller investor base and will consequently be under-priced. While impact investing may cause an upwards pressure in demand and therefore increase returns, the opposite strategy of excluding companies exhibiting irresponsible behavior could also cause excess returns as a consequence of exogenous demand shifts, which may cause the stock to become undervalued based on fundamental financial factors. Pastor, Stambaugh & Taylor (2019) show that agents’ tastes for ’green’ assets affect prices and that agents are willing to pay more for stocks from firms with a green profile, thereby lowering the firms’ costs of capital. They found that green assets have negative CAPM alphas, whereas brown assets have positive alphas and that agents who tilt their portfolios towards ’green’ assets and away from ’brown’ assets, earn lower expected returns.

Empirical evidence of how low demand creates excess returns includes Hong and Kaperczyk (2009), who used institutional investor behavior to show that

(11)

relatively lower demand for a portfolio of stocks from sinful industries, defined as companies from the tobacco-, alcohol- and gambling-industries, provided an annualized four-factor alpha of 3.7% from 1926 to 2006, relative to stocks from comparable industries.

An argument against the theory of negative demand shifts leading to increased returns is that other investors will tilt their portfolios in the opposite direction, expecting to earn quasi-arbitrage returns. Grossman and Stiglitz (1980), however, find that this strategy is too costly, and do not expect a fully offsetting effect. Shleifer (1997) theorize that one limit of arbitrage is that there may not be enough arbitrage capital available to offset large demand-shifts.

Literature on returns from ESG-investing in top financial journals is gen- erally sparse. Hartzman and Sussman (2019) analyzed fund flows and found that net inflows in socially responsible funds equated to $24 billion compared to net outflows of $12 billion in low-responsible funds, in the 10 months following the launch of the Morningstar Fund Sustainability Ranking in 2016, but did not find any subsequent difference in fund performance. Bebchuk, Cohen and Wang (2013) investigated known correlation between governance indices and abnormal returns, and found that the abnormal returns disappeared at the turn of the millennium.

ESG-ratings are marketed as a way to screen investment, potentially invoking behaviorally based asset demand shocks. In the traditional finance paradigm, demand shocks are absorbed by arbitrageurs, who can use sophisti- cated trading strategies to ensure that assets remain close to their equilibrium price. Theoretical work by De Long et al. (1990) and Shleifer and Vishny (1997) show how perfect arbitrage can break down, and empirical studies of the price effects of SP&500 listings (Harris and Gurel (1986); Beneish and Whaley (1996); Lynch and Mendenhall (1997)) provide compelling evidence of the importance of such breakdowns for the prices of individual stocks.

(12)

3 Methodology and Hypotheses

We start by testing for abnormal returns of value-weighted zero investment and long portfolios of the highest/lowest 10% stocks from 2010 to 2018, using Fama and French’s 5-factor model (2015) as our benchmark. ESG-ratings are posted in January, and are updated in regular intervals if new information becomes available. We choose to make the portfolio screening each June after publishing of the annual reports, holding the assets from July of year t to June of year t+ 1, with monthly rebalancing. We move on to tests correcting for cross-sectional correlation, using Fama and MacBeth’s (1973) test to measure the performance of portfolios of ESG-stocks in the presence of a series of known return predictors. We then investigate whether institutional owners own less/more equity in companies that have low/high ESG-ratings. Following Hong and Kacperczyk (2009), we apply a pooled panel OLS-model, controlling for several variables that are known to affect institutional investor behavior and preferences. Lastly, we perform several robustness checks, including running equal-weighted portfolios, testing several portfolio breakpoints, changing the formation month, analyzing outlier influence on return performance by changing winsorization level and netting out announcement returns to see how the performance is affected by earnings surprises.

When investigating the data we compute both time-series averages of Pear- son product-moment and non-parametric Spearman rank correlations for pair- wise variables in all of our linear models. Pearson product-moment correlations are computed using data sets winsorized at the 0.5% level and Spearman rank correlations are computed using our raw data sets² . If the Spearman rank correlation is substantially larger than the Pearson product-moment correlation for a given pair of variables, this likely indicates that there is a monotonic, non-linear relation between those variables. In the case of the opposite rela-

2Spearman rank correlations ranks the observations relative to other observations and winsorization distorts the rankings. Therefore, we use the raw data with Spearman rank correlation calculations.

(13)

tionship, it likely indicates that there are some extreme data points that are exerting strong influence on the calculations and that further winsorization is needed (Bali et al. (2016)). In general, with the exception of our time-series factor-regression, we do not worry too much about multicollinearity between our independent variables, since it does not distort the interpretation of our variables of interest. We limit ourselves to commenting on the lack of inter- pretability of coefficients for the correlated independent variables when deemed necessary.

In general, we take the logarithm of any variable if in our data inspection we recognise a significant improvement in reducing heteroskedasticity and non- normality of the residuals through visual inspection of bivariate plots.

3.1 ESG-Ratings and Returns

Empirical asset pricing often deals with portfolios of stocks rather than individual shares when explaining stock returns. In the literature, two main portfolio versions are used. The first is the Value-Weighted (VW) portfolio where all stocks are weighted according to their market capitalization at the time of portfolio formation. The return of a value-weighted portfolio pfor monthtis the sum of the weighted return of all portfolio assetsN, re-balanced on market capitalization each month from July of year tto June of year t+ 1.

Value-weighted portfolios put larger emphasis on the large market capitalization stocks in the portfolio. The Equal-Weighted (EW) portfolio gives every stock the same weight regardless of their market capitalization. These weight- ing strategies have very different risk strategies and practical implications.

Equal-weighted portfolios tend to be riskier as they tend to put a heavier emphasis on low-priced growth stocks. Additionally, value-weighted portfolios are more tax efficient, since re-balancing of an equal-weighted portfolio always entails selling the best performing stocks in the portfolio. We will therefore

(14)

apply value-weighted portfolios for the analysis, but will run equal-weighted portfolios as a robustness check.

3.1.1 ESG Portfolios

To analyze whether there is a relationship between ESG-ratings and risk- adjusted returns, we construct a theoretical portfolio long the 10% companies with the lowest ESG-ratings in yeart and short the 10% companies with the highest ESG-ratings in year t. We estimate the following model:

r_ESG_low,t−r_ESG_high,t = ˆα₀+ ˆβ_mkt(r_mkt,t−r_f,t) + ˆβ_{SM B}r_{SM B,t}+ ˆβ_{HM L}r_{HM L,t}+ βˆRM WrRM W,t+ ˆβCM ArCM A,t+ ˆt, t= 1, ..., T

(1) in addition to the long portfolio:

r_ESG_low,t−r_f,t = ˆα₀+ ˆβ_mkt(r_mkt,t−r_f,t) + ˆβ_{SM B}r_{SM B,t}+ ˆβ_{HM L}r_{HM L,t}+ βˆ_{RM W}r_{RM W,t}+ ˆβ_{CM A}r_{CM A,t}+ ˆ_t, t = 1, ..., T

(2)

We also run the same long portfolio with high-rated ESG companies. r_ESG_low,t is the the return of a portfolio of companies with low ESG-ratings in montht, r_ESG_high,t is a portfolio of companies with high ESG-ratings in monthtandr_f,t is the risk-free rate in montht. For brevity, we use the 10% level as our primary breakpoint, while also running a smaller sample of tests for percentile breakpoints 2.5%, 5% and 20% as robustness checks. We run portfolios with other breakpoints to address any concerns of our results stemming from an arbitrary percentile cutoff and to analyze the effects of changing portfolio breakpoints.

If there is a relationship between ESG-ratings and returns, we would expect to see larger alpha coefficients when we decrease the portfolio breakpoint to only include the best- and worst-performing stocks and a smaller alpha coefficient when when we expand the portfolio to include more stocks that lean towards a ’neutral’ rating.

(15)

Our benchmark model consists of the factors included in Fama and French’s (2015) 5-factor model, which is built upon the more famous Fama and French (1993) 3-factor model. The excess return on the market is denoted asM KT; the excess returns of small companies over big companies is denoted asSM B and the excess returns of high book-to-market stocks over small book-to- market stocks is denoted as HM L. The two newly added factors are RM W, defined as the excess returns of highly profitable companies versus low prof- itability companies andCM A, defined as the excess returns of firms that invest conservatively versus the firms that invest aggressively. In addition, we use the momentum-factorM OM in robustness checks, which is a fourth factor added by Carhart (1997) to Fama and French’s original 3-factors, a variable designed to capture the excess returns of stocks the top-performing stocks from the last 12 months over the returns from low-performing stocks. Our main coefficient of interest is the intercept α representing the excess return of the portfolios.

We calculate our standard errors using Newey and West’s (1987) autocorre- lation and heteroskedasticity robust standard errors with 4(T /100)^2/9 lags.³

3.1.2 Hypothesis 1

We hypothesize that returns for zero investment portfolios long stocks with low ESG-ratings and short stocks with high-ESG ratings should be significantly different from zero. Formally, the hypothesis is:

H₀: α₀ = 0 H₁: α₀ 6= 0

3Newey and West (1994) argue that the choice of lag length is arbitrary. We, nevertheless, choose our lag length based on the Bartlett’s kernel-specification of the given formula, which is widely used in econometric applications.

(16)

3.1.3 Cross-Sectional Regressions

We compare our results from the time-series regressions by running regressions based on the methodology created by Fama and MacBeth (1973), which aims to quantify the average reward for factor exposure. Unlike portfolio analysis, the Fama-Macbeth analysis allows us to control for a large set of other variables when examining the relation of interest. As the first step, we run monthly cross-sectional regressions for each month in our sample. This gives us slope coefficients on each independent variable for each period along with the associated standard errors for each month. To calculate the coefficients we take the means of the time-series coefficients. Fama and MacBeth then suggests that one should use the standard deviation of the cross-sectional regression estimates to generate the sampling errors for these estimates, but this approach has been widely criticized (e.g. by Cochrane (2009)), because we only have one sample mean for each cross-sectional regression, which ignores the cross-sectional estimation errors. An alternate approach, offered by Cuth- bertson (2004) is, instead of taking the standard errors of the sample mean, to take the mean of the standard errors. We will be using the latter in our calculations. Formally, we estimate:

r_i,t −r_f,t = ˆa₀+ ˆb₁ESGDU Mlow,t−1+ ˆb₂LOGSIZE1i,t−1+ ˆb₃BET A1i,t−1+ ˆb₄LOGM B1i,t−1+ ˆb₅RET ADJ1i,t−1+ ˆb₆AV GM RET1i,t−1+ ˆb₇LOGT U RN1i,t−1+ ˆb₈LOGAGEi,t−1+ ˆb₉BLEV1i,t−1+ ˆ_i,t, t= 1, ..., T, i= 1, ..., N

(3) wherer_i,t−r_f,t is the excess return on asset i at time t, b₁ is our coefficient of interest, whereESGDU Mlow,t−1 is a dummy variable, which equals one if the company had an ESG-rating among the bottom 10% in month t−1, based on scores from July of year t to June of year t+ 1, and zero otherwise. We

(17)

run the same specification for the high-rated ESG portfolio as well. All of our independent variables are lagged by one month and consist of a series of variables that have been found to be predictors of abnormal returns. If the Efficient Market Hypothesis (Fama (1970)) holds, then all of our independent variables should be statistically indistinguishable from zero. An explanatory list of variables, including its technical construction and article source can be found in Appendix A.

3.1.4 Hypothesis 2

We hypothesize that dummy coefficients of portfolios consisting of companies with low or high ESG-scores are significantly different from zero when controlling for the presence of a series of known return predictors. Formally, the hypothesis is:

H₀: b₁ = 0 H₁: b₁ 6= 0

3.2 Institutional Ownership Regressions

To empirically test whether institutions such as pension funds, universities, religious organizations, banks, and insurance companies perform impact investing or negative screening, we develop a model based on methodology from Hong and Kacperczyk (2009). We estimate several permutations of the following panel OLS regression:

(18)

IO_i,t = ˆc₀+ ˆd₁ESGDU M_i,t + ˆd₂LOGSIZE_i,t + ˆd₃BET A_i,t+ ˆd₄LOGM B_i,t+ dˆ₅LOGY IELD_i,t+ ˆd₆LOGAGE_i,t+ ˆd₇LOGP RIN V_i,t+

dˆ₈LOGST DRET_i,t+ ˆd₉AV GM RET_i,t+ ˆd₁₀LOGBB_i,t+ dˆ₁₁LOGT U RN_i,t+ ˆd₁₂N ASDAQ_i,t+ ˆd₁₃S&P500_i,t+ ˆ_i,t, i= 1, ..., N

(4) whereIO_i,tis the percentage of ownership for companyiat timetandd₁is our coefficient of interest which measures whether stocks in pre-defined low/high- ESG rated portfolios have different level of ownership than other stocks. Our other control variables are based on extensive research of institutional investor behavior and aim to control for a complete set of factors that explain institutional investors’ investment patterns. These control variables can broadly be divided into four different categories, where institutional investors according to literature have preferences based on liquidity and transaction cost motives, prefer less volatility, stocks that are predicted to do well given known return anomalies and stocks with different payout structures. Our main purpose is to soak up as much of the cross-sectional variation as possible so that the regression results purely reflect the difference in ownership for our variable of interest. An explanatory list of variables, including its technical construction and article source can be found in Appendix A.

To address the concern of regression standard errors, conditional on the independent variables, are clustered within groups of industries, we use Moul- ton’s clustered standard errors (1986) on Fama and French’s 48-industries (1997), following Hong and Kacperczyk (2009). If standard errors are clustered among industries, this causes a loss in the precision of the estimators, and Moulton show that one can correct these estimates by imposing an infla- tion term on the standard errors given by:

(19)

τ_j '1 +ρ_xjρ_u

V[N_g]

N_g

+N_g−1

(5) where ρxj is a measure of the within-cluster correlation of xj, ρu is the within cluster error-correlation,Ng is the correlation of clusterg and Ng is the average cluster size.

3.2.1 Hypothesis 3

We expect that institutional investors, on average, reacts to ESG-ratings and invests significantly more or less in stocks with high or low ESG-ratings and hypothesize:

H₀: d₁ = 0 H₁: d₁ 6= 0

4 Data

This section is divided into three parts. We first describe the databases and merging procedures, then describe our screening and cleaning methodology.

The last part is a brief description of descriptive statistics and correlations for our data sets.

4.1 Databases and Merging

We get market data from the Center for Research in Security Prices (CRSP) and fundamental accounting data from COMPUSTAT. We apply CRSP’s permno as our primary security identifier. To match the two databases we use the CRSP/COMPUSTAT merged database. We get data for institutional ownership from Thompson Reuters’ 13-F database. The ESG-scores are retrieved from Refinitiv Eikon, but the database share no common identifier with CRSP or COMPUSTAT data, so we perform several name and ticker string matching techniques along with manual matching to link the data via

(20)

the cusip-identifiers of Institutional Brokers Estimate Systems (IBES), which serves as a bridge between Refinitiv and CRSP⁴ . To get accurate daily data for S&P500 listings we use COMPUSTAT’s Index Constituents database. The data for dividend yield has been retreieved from WRDS’ Financial Ratios Suite.

4.2 Data Screening and Cleaning

We employ similar screening procedures to those of Fama and French (1992). We exclude financial firms, defined as those starting with a one-digit sic code of 6, because the leverage level is incomparable with companies from other industries. We also exclude companies in July of year t if it is missing a stock price in CRSP for either December of year t −1 or from June of year t. Companies missing monthly returns data for more than 36 out of the last 60 months are also excluded along with firms with missing or negative book equity values in COMPUSTAT. We only analyze assets classified as common stocks (CRSP shrcd must be 10 or 11) and shares must be listed on the New York Stock Exchange (NYSE), NASDAQ or the American Stock Exchange (AMEX) (CRSP exchcd-variable must be 1, 2 or 3). All daily and monthly returns are adjusted using data from theCRSP Stock Events - Delist- ing Information-database. This database takes into account realized returns for investors who held firms during events such as bankruptcies or takeovers, where this was not reflected in the listed stock price.⁵ Whenever we take the logarithm of a variable with a portion of logically explainable zero-values we add a constant to all variables in our sample to not erroneously discard valid observations. We winsorize all data on the 0.5% level, with the exception of data retrieved from Kenneth French’s Data Library (KFDL).

4Code for linking CRSP and IBES with Python along with several open source code sections that we’ve used as inspiration is available at WRDS: https://wrds- www.wharton.upenn.edu/pages/support/applications/python-replications/

5The CRSP-Delisting database has been accused of inaccuracies and incomplete data (Shumway (1997)), but besides pointing this out here, we do not address this further in our analysis.

(21)

4.3 Descriptive Statistics and Correlations

We here report the descriptive statistics and correlations from the period spanning from 2010 to 2018 for the value-weighted zero investment and two long portfolios along with the Fama and French’s 5 factors (2015). Both the median and mean returns for the zero investment portfolio are positive, indicating that the portfolio of low-rated ESG stocks have outperformed high- rated ESG-stocks, before adjusting for risk. The long portfolios show that the standard deviation is lower for the low-rated portfolio, which means that the value-weighted long portfolio of companies with low ESG-ratings also returned a higher Sharpe Ratio in the period. Correlations reveal that the zero investment portfolio shows low to moderate positive correlation with the long low-ESG portfolio and is similarly negatively correlated with the high-ESG portfolio. The long portfolios both show moderate to high correlation with the market factor. Descriptive statistics and correlations for our different regression data sets are reported in Appendix E. One significant point of note is that the mean holdings of institutional investors rose from 39.1% in the sample running from 1980 to 2018, while our sample of interest saw it increase to 61.6%. This implies that institutional investors have become an increasingly dominant investor class over the past decade.

(22)

Table1:DescriptiveStatisticsandCorrelationsTime-SeriesRegressions2010-2018 Thistablereportsdescriptivestatisticsandcorrelationsforthetime-seriesfactorregressionsfromJulyof2010throughout2018.Wereportthetotalnumberofobservations(N),theminimum value(Min),the5thpercentile(5th),the25thpercentile(25th),themedian(Median),the75thpercentile(75th),the95thpercentile(95th),themaximumvalue(Max),thestandarddeviation (Std),theskew(Skew)andtheexcesskurtosis(Kurt).ThetophalfofthetablereportsdescriptivestatisticsforourmainportfoliosofinterestandtheFama-French5-factors.Thelowerhalf ofthetablereportsaveragesoftheannualcross-sectionalPearsonproduct-momentandSpearmanrank-correlationsbetweenpairsoffactors.Below-diagonalentriespresenttheaveragePearson product-momentcorrelations.AbovediagonalentriespresenttheaverageSpearmanrankcorrelation.ESGlow−ESGhighisavalue-weightedzeroinvestmentportfoliolongthe10%highestrated ESG-stocksandshortthehighestratedESG-stocks.ESGhigh−Rfisthereturnfromavalue-weightedportfoliolongthe10%highestrated-ESGstocksandESGlow−Rfisthereturnfroma value-weightedportfoliolongthe10%lowest-ratedESGstocks.MKTistheexcessmarketreturn,withRFbeingtheonemontht-billrate.SMBistheaveragereturnofninesmallstockportfolios minustheaveragereturnonninebigstockportfolios.HMListheaveragereturnoftwovalueportfoliosminustwogrowthportfolios.RMWistheaveragereturnontworobustoperatingportfolio minustheaveragereturnontwoweakoperatingprofitportfolios.CMAistheaveragereturnontwoconservativeinvestmentportfoliosminustheaveragereturnontwoaggressiveinvestmentportfolios. DescriptionNMin5th25thMedian75th95thMaxMeanStdSkewKurt ESGlow−ESGhigh102-0.046-0.034-0.0120.0040.0160.0460.0660.0030.0230.4580.254 ESGhigh−Rf102-0.090-0.052-0.0090.0120.0300.0550.1170.0090.037-0.0901.082 ESGlow−Rf102-0.080-0.046-0.0110.0160.0350.0630.0890.0120.033-0.201-0.099 MKT102-0.096-0.060-0.0070.0110.0310.0680.1140.0110.036-0.3000.823 SMB102-0.046-0.038-0.0190.0010.0130.0360.068-0.0000.0230.179-0.273 HML102-0.041-0.032-0.015-0.0030.0090.0360.083-0.0020.2130.8841.526 RMW102-0.040-0.021-0.0100.0020.0010.0260.0350.0010.015-0.130-0.215 CMA102-0.033-0.022-0.010-0.0000.0090.0240.0370.0000.0140.223-0.207 CorrelationsESGLMH,10%ESGhigh10%ESGlow10%MKTSMBHMLRMWCMA ESGlow−ESGhigh1-0.4600.290-0.087-0.033-0.240-0.072-0.225 ESGhigh−Rf-0.40510.5840.7520.2150.183-0.2160.184 ESGlow−Rf0.2600.72910.7310.2870.000-0.1950.088 MKT-0.1700.8940.77210.2850.049-0.346-0.029 SMB0.0270.3600.4160.34110.017-0.511-0.028 HML-0.0440.1290.1060.0840.1031-0.1760.559 RMW-0.013-0.200-0.231-0.290-0.475-0.23910.012 CMA-0.1230.0580.031-0.056-0.0100.6300.0461

(23)

5 Results

5.1 Time-Series Factor Regression Results

Table 2 shows the results of 5-factor time-series regressions for a value- weighted portfolio long the 10% bottom-rated and short the 10% top-rated stocks between 2010 and 2018, along with their separate long portfolios. Our results are consistent with the findings of Pastor, Stambaugh & Taylor (2019), who found that investors who prefer responsible assets, earn lower expected returns. All specifications for the zero investment portfolio are statistically significant at the 1% level, with a stable intercept, culminating in a 5-factor alpha of 56 basis points. The different long portfolios show that the largest influence comes from strong performance from the low-rated ESG portfolio. The M KT-coefficient for the low-rated portfolio is considerably lower, indicating that it carries less systematic risk. The CM A-coefficient indicates that the high-rated ESG firms invest more conservatively, significant at the 1% level.

Table 2: Time-Series Regressions - 10^th percentile

Results from time-series regressions of value-weighted portfolios. ESGlow−ESGhigh is a portfolio long the bottom-rated 10% ESG companies and short the 10% highest-rated ESG companies from 2010-2018.

ESGlow−Rf is a portfolio long the 10% lowest-rated, and ESGhigh−Rf is a portfolio long the 10%

highest-rated ESG-companies. Portfolio composition is changed in June of each year. M KT is the market premium.SM Bis the return of a portfolio long small stocks and short large stocks. HM Lis the return of a portfolio long high book-to-market stocks and short low book-to-market stocks,RM W is the return of a portfolio long the most profitable companies and short the least profitable companies.CM Ais the returns of a portfolio long conservative investment companies and short aggressive investment companies. Standard errors are adjusted for serial correlations using Newey West (1987) standard errors. ∗∗∗1% significance;

∗∗5% significance;∗10% significance.

2010-2018, value-weighted ALPHA MKT SMB HML RMW CMA ESG_low−ESG_high 0.0052∗∗∗ -0.1837∗∗

(0.002) (0.079)

ESGlow−ESGhigh 0.0054∗∗∗ -0.2009∗∗ 0.0692 (0.002) (0.084) (0.126)

ESG_low−ESG_high 0.0054∗∗∗ -0.2011∗∗ 0.0724 -0.0220 (0.002) (0.084) (0.123) (0.122)

ESG_low−ESG_high 0.0053∗∗∗ -0.2000∗∗ 0.0757 -0.0210 0.0135 (0.002) (0.083) (0.152) (0.122) (0.176)

ESGlow−ESGhigh 0.0056∗∗∗ -0.2048∗∗ 0.0832 0.1029 0.0518 -0.2876 (0.002) (0.082) (0.150) (0.158) (0.169) (0.207) 2010-2018, value-weighted ALPHA MKT SMB HML RMW CMA ESGlow−Rf 0.0040∗∗ 0.7594∗∗∗ 0.2021∗ 0.0416 0.2998∗∗ 0.0794

(0.002) (0.067) (0.110) (0.108) (0.121) (0.166) ESGhigh−Rf -0.0016 0.9642∗∗∗ 0.1189∗ -0.0612 0.2480∗∗ 0.3670∗∗∗

(0.001) (0.037) (0.069) (0.066) (0.113) (0.109)

(24)

5.2 Fama MacBeth Regression Results

Since results show that over-performance from low-rated firms is greater than the under-performance from high-rated ESG firms, we here continue with a primary focus on low-rated ESG firms (see Appendix Table 16 for a similar specification with the high-rated ESG-portfolio). Table 3 presents parameters that are estimated using the Fama-MacBeth (1973) method, with standard errors using specifications by Newey and West (1987). The dependent variable is the excess return on stock i in period t, and the variable of interest is ESGDU M_low, which is a dummy variable equal to one if the company has an ESG-rating in the bottom 10%, and zero otherwise. We add variables one by one to see the effects of the variables on the dummy coefficient. Statistical significance of independent variables indicates that these had some predictive power on future returns in the regression period. RET ADJ1 is the one-month momentum factor, and is negative and statistically significant at the 1% level, consistent with the findings of Jegadeesh (1990), who showed that stocks tend to exhibit short-term momentum reversal. AV GM RET1 is the rolling 12- month average return, and is positive and statistically significant at 1% for all specifications with the exception of the last, where it remains significant at the 5%-level. This is consistent with the findings of Jegadeesh and Titman (1993), who showed that past winners had a tendency to continue to do well and past losers had a tendency to keep under-performing. LOGT U RN1 is negative and significant at the 1% level, consistent with theilliquidity premium(e.g Stoll and Whaley (1983)). The size coefficient, denoted byLOGSIZE1, is the variable that impose most influence on our dummy coefficient of interest. Inconsistent with the findings of Fama and French (1993), who showed that small companies have had a tendency to outperform large companies, the coefficient is positive and significant at the 1% level, reducing the size of the portfolio-coefficient from 0.0047 to 0.0033. While the coefficient remains significant at the 5%

(25)

level, this drop in magnitude indicates that the over-performance from the value-weighted portfolio consisting of low-rated ESG firms is in part driven by a positive contribution from large firms.⁶

Table 3: Fama Macbeth - Company, 10% ESG 2010-2018

This table reports results from Fama and Macbeth (1973) cross-sectional regressions for the period 2010-2018 on the monthly return of stocks net of the risk-free rate on the lagged values of a set of well-known predictors of stock returns.ESGDU Mlowis a dummy variable which equals one if the company has an ESG-rating amongs the bottom 10% in yeart, with ranking being registered starting from July each year. BET A1 is the 36-month rolling company beta. LOGM B1 is the logarithm of the market-book ratio.RET ADJ1 is the monthly return of the company adjusted for delisting returns. AV GM RET1 is the average 12-month return. LOGT U RN1 is the logarithm of average daily share turnover, during the past year.LOGAGE is the logarithm of the age of the company. BLEV1 is the book-leverage of the company.

LOGSIZE1 is the logarithm of the market capitalization. Standard errors are adjusted for serial correlation using standard errors as in Newey and West (1987).∗∗∗1% significance;∗∗5% significance;∗10% significance.

2010-2018 (1) (2) (3) (4) (5) (6) (7) (8)

ESGDU Mlow 0.0054∗∗∗ 0.0053∗∗∗ 0.0052∗∗∗ 0.0050∗∗∗ 0.0052∗∗∗ 0.0047∗∗∗ 0.0047∗∗∗ 0.0033∗∗

(0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002)

BETA1 -0.0018 -0.0019 -0.0022 -0.0022 -0.0014 -0.0013 -0.0013 -0.0012

(0.002) (0.002) (0.002) (0.002) (0.001) (0.001) (0.001) (0.001)

LOGMB1 0.0008 0.0009 -0.0002 0.0002 0.0004 0.0005 -0.0007

(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001))

RETADJ1 -0.0149∗∗ -0.0206∗∗∗ -0.0203∗∗∗ -0.0209∗∗∗ -0.0209∗∗∗ -0.0221∗∗∗

(0.007) (0.006) (0.006) (0.006) (0.006) (0.006)

AVGMRET1 0.0884∗∗∗ 0.0860∗∗∗ 0.0833∗∗∗ 0.0807∗∗∗ 0.0703∗∗

(0.031) (0.031) (0.030) (0.030) (0.029)

LOGTURN1 -0.0034∗∗∗ -0.0033∗∗∗ -0.0033∗∗∗ -0.0047∗∗∗

(0.001) (0.001) (0.001) (0.001)

LOGAGE 0.0023∗∗ 0.0023∗∗ 0.0006

(0.001) (0.001) (0.001)

BLEV1 0.0006 -0.0022

(0.003) (0.003)

LOGSIZE1 0.0019∗∗∗

(0.001)

5.3 ESG-ratings and Institutional Ownership

We follow the approach proposed by Hong and Kaperczyk (2009) of running a pooled panel OLS regression with Moulton’s (1986) standard errors clustered at the 48-industry level, with institutional ownership as the dependent variable. The ESGDU M_low-variable is defined similarly as in section 5.2. If the difference in performance is related to active investment strategies from institutional investors, such as negative screening or impact investing, this should be reflected by the coefficient being significantly different after

6See Appendix for time-series regressions with equal-weighted portfolios.

(26)

controlling for other factors proven to influence their investment behavior.

Our results are reported in Table 4. Specification 1-5 are different permutations of our independent variables, where moderately correlated variables are rotated. In the 6^th we include all variables, except for LOGP RIN V, since the price inverse shows strong, negative correlation with LOGSIZE, and LOGY IELD, which is a consequence of a non-linear preference for dividend yield from institutional investors which makes interpretation difficult⁷. If institutional investors on average have performed negative screening of the 10% lowest-rated ESG stocks, we expect to see a negative and statistically significantESGDU M_low-coefficient, yet the coefficient is consistently positive.

Only two specifications have statistically significantESGDU M_low coefficients, but they both have a positive coefficient sign, indicating that institutional owners hold more low-rated ESG stocks. The strongest result is from permutation 3, which has a size of 0.0474 and is statistically significant at the 1%

level, but the permutation does not control for firm size. A coefficient size of 0.0174 in our 6^thspecification indicates that institutional investors hold 1.74%

more stocks in low-rated ESG firms in absolute terms, and approximately 2.8%

more in relative terms, which is of little economic significance even if it had been statistically significant. When splitting the dependent variables into sub- groups, where regression specification 7 refers to holdings by banks, insurance companies and ’other’ institutional owners, and specification 8 refers to stock ownership by mutual funds and independent investment advisors, the former group holds 2.71% more stocks in the low-rated ESG-firms, statistically significant at the 5% level. These investors also tend to hold significantly less momentum stocks, and significantly more of high trading volume stocks. This is consistent with a focus on long-term investing and low-cost trading strategies, which are both somewhat inconsistent with trading strategies related to

7Grinstein and Michaely (2005) found that institutional investors prefer companies that pay dividend yield, but prefers companies that pay low dividends yield over companies that pay high dividend yields

(27)

Table 4: Institutional Ownership, 10% Lowest Rated

This table reports summary statistics for the variables used for the eight sets of regressions. In the first six, the dependent variable is overall institutional ownership (IO), which is calculated at the end of each year. In regression (7) the dependent variable is the aggregate ownership of Thompson Reuters category owners (1),(2) and (5); banks, insurance companies and other. In regression (8) the dependent variable is owner types (3) and (4); mutual funds and independent investment advisors.ESGDU Mlowequals one if a stock is amongst the 10% lowest rated ESG-companies and zero otherwise. LOGSIZE is the logarithm of the market capitalization of the company. BETAis the firms industry beta. LOGMB is the logarithm of the market-to-book ratio. LOGYIELD is the logarithm of the yearly dividend ratio divided by the price at the end of the year. LOGAGEis the logarithm of the number of years the company has been listed at COMPUSTAT at the end of the year. LOGPRINV is the logarithm of the inverse of the price at the end of the year. LOGSTDRET is the daily stock return standard deviation during the past year. AVGMRET is the average monthly return during the past year. LOGBB is the logarithm of the buyback ratio of the company during the past year.LOGTURN is the logarithm of average daily share turnover during the past year.NASDAQequals one if the company is listed on NASDAQ and zero otherwise. S&P500 equals one if the company is on the S&P500-index and zero otherwise. These are the results of pooled OLS regressions with Moulton’s (1986) standard errors, clustered at the 48-industry groupings. The ownership data covers the period 2010-2018.∗∗∗1% significance;∗∗5% significance;∗10% significance.

Variable (1) (2) (3) (4) (5) (6) (7) (8)

ESGDU Mlow 0.0169 0.0282∗ 0.0474∗∗∗ 0.0218 0.0239 0.0174 0.0271∗∗ -0.0106 (0.019) (0.016) (0.016) (0.019) (0.019) (0.016) (0.013) (0.008) LOGSIZE 0.1263∗∗∗ 0.1317∗∗∗ 0.0863∗∗∗ 0.0916∗∗∗ 0.0651∗∗∗ 0.0258∗∗∗

(0.005) (0.006) (0.007) (0.012) (0.008) (0.004)

BETA 0.1324∗∗∗ 0.0717∗∗ 0.1239∗∗ 0.1576∗∗∗ 0.1236∗∗ 0.1331∗∗∗ 0.0785∗∗ 0.0464∗∗∗

(0.042) (0.035) (0.053) (0.044) (0.050) (0.048) (0.034) (0.016) LOGMB 0.0027 0.0154∗ -0.0178∗∗∗ -0.0162∗∗ -0.0154∗∗∗ -0.0005

(0.007) (0.008) (0.006) (0.007) (0.005) (0.002)

LOGYIELD -0.0078∗∗∗ -0.0073∗∗∗

(0.002) (0.002)

LOGAGE -0.0022 -0.0005 -0.0072 -0.0000 -0.0075∗∗∗

(0.007) (0.006) (0.006) (0.004) (0.003)

LOGPRINV -0.1574∗∗∗

(0.005)

LOGSTDRET -0.3914∗∗∗ -0.1443∗∗∗ -0.0954∗∗∗ -0.0507∗∗∗

(0.025) (0.027) (0.020) (0.008)

AVGMRET -0.0461 -0.2264∗∗∗ -0.1999∗∗∗ -0.0146

(0.057) (0.067) (0.052) (0.021) LOGBB 0.0068∗∗∗ 0.0098∗∗∗ 0.0070∗∗∗ 0.0054∗∗∗ 0.0015∗∗∗

(0.002) (0.002) (0.002) (0.001) (0.001)

LOGTURN 0.1759∗∗∗ 0.0699∗∗∗ 0.0957∗∗∗ 0.0712∗∗∗ 0.0238∗∗∗

(0.006) (0.011) (0.013) (0.009) (0.004)

NASDAQ 0.0079 0.0069 0.0100 0.0072 0.0028

(0.013) (0.013) (0.015) (0.011) (0.004)

S&P500 -0.2892∗∗∗ -0.0597∗∗∗ -0.0923∗∗∗ -0.2845∗∗∗ -0.2664∗∗∗ -0.1922∗∗∗ -0.0717∗∗∗

(0.022) (0.011) (0.015) (0.023) (0.028) (0.021) (0.008)

short- to medium-term fluctuations of ESG-ratings. While theESGDU M_low- coefficient is statistically significant for some permutations, it appears difficult to reject the null hypothesis of zero difference between institutional ownership of low-rated ESG stocks and other stocks. Running a similar regression with a dummy variable consisting of the 10% highest-rated ESG-firms yields similar results (Appendix Table 18).

(28)

5.4 Institutional Ownership of Industries

While we could not find evidence of negative screening of single stocks, there could potentially be more stigma tied to investing in low-rated ESG industries, such as investing in industries known for high levels of carbon emissions. To test for this, we select the industries with the lowest ESG-ratings using time- series means from 2010 to 2018⁸. We follow Hong and Kaperczyk’s (2009) approach of investigating how well these companies perform against a portfolio of comparable industries. The lowest-rated industries are Soda, Coal, Fun, Tobacco and Fabricated Products, where our chosen comparisons are Beer, Oil, Toys, Food and Steel. Hong and Kaperczyk constructed a dummy variable equal to one if a company resides in either of the low-rated industries or comparable industries (GDU M). We include GDU M_low to separate between institutional ownership differences that are caused by ESG-scores from ownership differences caused by unrelated trends⁹. Table 5 reports the overall results, where ESGIN DDU M_low is large, negative and significant on the 5% level for all regression specifications, except for a regression specification 5, in which we do not control for the preferences of companies listed on the S&P500-index. A negative coefficient of -0.0851 in our last regression specification indicates that institutional owners held approximately 14% less of the market cap in low rated ESG industries compared to other industries, after controlling for known investment preferences. Investigations from preceding time-periods shows a marked drop in institutional ownership for these industries around 2010, consistent with negative screening caused by an increased focus on ESG-ratings (Appendix Table 19).

One concern when performing this analysis, however, is that our results could be heavily influenced by the ’sin’-effect, proposed by Hong and Kaper-

8see Appendix B for a full list of industries and time-series mean ratings

9E.g. GDU Mlowis able to make a distinction between differences in divestment from coal that are caused by low ESG-ratings and a general trend of divesting in stocks from companies involved in fossil fuel industries, because the dummy variable also includes companies from the comparable oil-industry

ESG-Ratings and Returns

GRA 19703

Master Thesis

ESG-Ratings and Returns

Master Thesis

Acknowledgements

Contents

List of Abbreviations

List of Tables

List of Symbols

1 Introduction

2 Literature Review

3 Methodology and Hypotheses

3.1 ESG-Ratings and Returns

3.2 Institutional Ownership Regressions

4 Data

4.1 Databases and Merging

4.2 Data Screening and Cleaning

4.3 Descriptive Statistics and Correlations

5 Results

5.1 Time-Series Factor Regression Results

5.2 Fama MacBeth Regression Results

5.3 ESG-ratings and Institutional Ownership

5.4 Institutional Ownership of Industries